# AI that earns trust

> Users don’t need magic. They need predictable failure modes, clear ownership of answers, and products that know when to shut up.

- Author: Kevin Dalvi
- Published: 2026-06-01
- Read time: 8 minutes
- Tags: ai, product, trust
- Canonical: https://www.kevindalvi.com/blog/ai-that-earns-trust

I’ve shipped AI features that impressed in demos and disappointed by Tuesday. The pattern was always the same: we optimized for wow, not for recovery. Trust isn’t built when the model is right. It’s built when the product behaves well when the model is wrong.

Users forgive mistakes they can see. They don’t forgive confident nonsense dressed as expertise.

## Design for visible uncertainty

I treat model output as a draft with a source trail, not as UI chrome that asserts truth. Citations, ‘why we think this,’ and easy edit paths matter more than prettier completion animations. If the system can’t show its work, it shouldn’t speak in a brand voice that sounds sure.

When we can’t cite a source, we say so. Silence beats a hallucinated footnote. Users learn faster from a product that admits gaps than from one that papers them over with fluency.

- Label AI generated content in the interface. Always
- Prefer suggest and confirm over silent automation on irreversible actions
- Log refusals and low confidence paths as first class product events
- Give users a one click path to human override or plain tools

> A trustworthy AI product is mostly a boring product with a clever engine. Not the reverse.

## Scope is a trust feature

The fastest way to burn credibility is a general assistant bolted onto a vertical problem. I narrow the job until evaluation is possible. Can we measure success without a vibe check? If not, the feature isn’t ready. The model might be.

Brand shows up here too. Warm copy can’t cover opaque behavior. I’d rather a product sound plain and precise than witty and slippery. Personality is earned after reliability.

I also instrument disagreement. When users edit or discard a suggestion, that’s gold. Not a failed completion. Products that only celebrate accept rates train themselves to be agreeable, not useful.

## Earn the next permission

Trust compounds like credit. Start with low stakes assists. Graduate to semi automated flows only after users have edited, rejected, and survived a failure with you. Each permission should feel like a promotion the product deserved.

I’m not interested in AI that feels magical for five minutes. I’m interested in AI that still has a job after the novelty tax is paid. That’s a product discipline problem. And it’s the only kind worth solving.